This Article Abstract Figures Only Full Text (PDF) Services Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Astley, S M Search for Related Content PubMed PubMed Citation Articles by Astley, S M

Computer-based detection and prompting of mammographic abnormalities

Imaging Science and Biomedical Engineering, Stopford Building, Oxford Road, Manchester M13 9PT, UK

	Abstract

Top Abstract Introduction Computer-based detection of... Computer-based aids for film... The potential of CAD... The clinical evaluation of... Conclusions References

 
Mammographic film reading is a highly demanding task, particularly in screening programmes where the reader must perform a detailed visual search of a large number of images for early signs of abnormality, which are often subtle or small, and which occur very infrequently. False negative cases, where signs of abnormality are missed by a film reader, are known to occur. Computer based algorithms can be used to detect abnormal patterns in images, but it is not possible to reliably detect all signs of abnormality in mammograms, so screening cannot yet be fully automated. The most successful detection algorithms are, however, incorporated in computer-aided detection (CAD) systems which indicate potentially abnormal locations to the reader in a process known as prompting. CAD systems have the capacity to reduce the frequency of false negative errors by ensuring that suspicious regions of the images are thoroughly searched and by increasing the weighting attached to subtle signs that may otherwise have been dismissed. One of the areas currently being researched is the effect of prompting on human performance. This is complex, since readers are presented with prompts generated by multiple detection algorithms, each of which has a different sensitivity and specificity. This paper reviews progress in abnormality detection, the strengths and the weaknesses of CAD, and the methodologies used to evaluate CAD in a clinical setting.

	Introduction

Top Abstract Introduction Computer-based detection of... Computer-based aids for film... The potential of CAD... The clinical evaluation of... Conclusions References

 
X-ray mammography has been shown to be effective as a method for detecting early breast cancer, but the success of mass screening depends critically on the availability of highly skilled film readers to interpret the images. The majority of film readers in the UK are consultant radiologists, although more recently breast physicians and experienced breast radiographers have also been trained to interpret mammograms. In order to maintain a sufficiently high standard of interpretation, readers are required to undergo training, to keep in practice and to evaluate their performance at regular intervals [1].

Mammographic film reading is a particularly demanding visual task. In screening programmes, the film reader must search for extremely infrequent and often very subtle signs of cancer superimposed on complex and variable backgrounds. Early breast cancer may appear in a variety of forms: a few particles of microcalcification; a small ill-defined or spiculated mass; abnormal asymmetry between right and left breast images, or subtle distortion of the underlying structure of the breast. These abnormalities vary in size, shape, structure, brightness and location and may share a great deal of similarity with normal mammographic appearances.

False negative cases, in which signs of cancer are missed by a reader, sometimes occur. Retrospective evaluation of the previous screening films of cancers detected between screening rounds (interval cancers) and screen-detected cancers show evidence of abnormality in between 16% and 27% of cases. Some of these signs are very subtle, and may have been seen by the readers but dismissed as being insignificant, but others are clear signs of malignancy [2–4]. However, different readers miss different cancers, as is evidenced by the success of double reading in which two readers independently read the films [5]. The most accurate method of interpretation is double reading with arbitration, where a third reader reviews cases about which the two readers disagree [5, 6].

In the UK, recent extensions to the National Health Service Breast Screening Programme (NHSBSP) have included increasing the age range of women invited for screening, and taking a second radiographic projection of each breast at all visits. These, coupled with a natural expansion of the eligible population, have significantly increased the manpower requirements of the programme. Different methods of coping with this, including the use of computer-based aids, are currently being explored.

Researchers have been developing algorithms to detect mammographic abnormalities for more than 30 years with the aim of either automating mammographic interpretation or, more realistically, providing a tool which will enhance human film-reading performance. A benchmark against which the efficacy of such tools can be measured is whether or not the performance of an individual reader could be improved to the extent that double reading was no longer necessary, as this would alleviate any shortage of film readers.

	Computer-based detection of mammographic abnormalities

Top Abstract Introduction Computer-based detection of... Computer-based aids for film... The potential of CAD... The clinical evaluation of... Conclusions References

 
There are two basic approaches to the problem of detecting abnormalities in mammograms: either to search the images for specific appearances suggestive of cancer, or to characterize normal mammographic appearance to the extent that it is possible to detect anything that fails to conform to the generated model of normality. Most published methods adopt the first strategy, although there is increasing interest in novelty detection in biomedical images and signals [7–9]. The main difficulty with this approach in mammography is the wide spectrum of the normal appearance in which any abnormalities are embedded. In addition, the quantity and pattern of glandular tissue alter both as a woman ages and in response to other hormonal influences such as hormone replacement therapy (HRT). A normal appearance in one mammogram might look highly suspicious in another, and differences between left and right breasts or unexpected temporal changes might also be clinically significant. The alternative approach, based on searching the images for specific patterns associated with different types of abnormality, has been more widely investigated, and algorithms targeted at several different signs of cancer have been developed and applied with varying degrees of success.

Microcalcifications are sometimes difficult for the human film reader to detect because of their small size and low contrast, particularly if they are superimposed on dense glandular tissue. However, of all the signs of abnormality found on mammograms, microcalcifications are the easiest to detect automatically. Unlike small ill-defined masses, which may superficially resemble normal glandular tissue, microcalcifications have properties – namely their very small size and high attenuation – which differ significantly from those of normal background structures. Computers can be trained relatively easily to detect small, bright, regions with well-defined edges (Figure 1). Most microcalcification detection algorithms make an initial attempt at detection based on these properties, followed by a more detailed analysis to reduce the number of false positives due to benign calcification, overlapping narrow linear structures, or artefacts such as screen–film "shot" noise [10]. A wide variety of detection methods have been investigated including mathematical morphology [11, 12], the use of matched filters [13] and neural networks [14, 15]. One successful approach involved a preliminary phase of noise estimation and equalization followed by a method based on Bayesian techniques and a Markov random field model. Iterative updating maximized the probability of a correct labelling of pixels [16]. It achieved more than 90% sensitivity with slightly less than one false detection per image on a set of 40 mammograms; derivatives of this method are frequently used as a basis against which to assess the performance of more recent techniques. The false positive removal (or analysis) phase of detection algorithms not only takes into account the shape, size and brightness of individual candidate particles, but also involves assessment of number, location and clustering of candidates; for example, isolated candidates can generally be ignored.

View larger version (2K): [in this window] [in a new window]   Figure 1. A cluster of microcalcification particles detected using a morphological detection algorithm [47].

Soft tissue masses are more difficult to detect automatically than microcalcifications because they may superficially resemble normal background structures such as overlapping areas of glandular tissue. Edge detection, which is one of the most widely used computer-based feature detection techniques, is of little use since the lesion edges are often not as strong as those of normal structures within the breast. In fatty breasts, methods directed at local increases in density of a particular size are reasonably effective, but in breasts with a greater glandular component many false positives result, and the problem again becomes one of analysis to determine which detected regions correspond to genuine abnormalities. An alternative approach is to look directly for the foci of radiating patterns of lines characteristic of spiculated lesions [22, 23]. The resulting locations can be compared with local increases in density to provide additional confirmation that a malignant lesion may be present. Linear structure analysis can also be used to detect architectural distortion, along with methods which first detect the outline of the glandular disc and then search around it for regions with an uncharacteristic shape [24]. The segmentation of the glandular disc has been tackled by various methods including texture measures and calibration of densities [25–27]. Mass detection algorithms are generally less sensitive and specific than microcalcification detection algorithms; commercial computer aided detection (CAD) systems claim sensitivities of up to 98.5% for microcalcifications, but only 89.2% for masses and distortions [28–30].

Both masses and diffuse asymmetries can be detected by comparing right and left breast images. Normal mammograms usually show some degree of asymmetry (Figure 2), but although this is usually easily identified and ignored by human readers, it can be problematic for computer-based approaches. The automated detection of asymmetry is a particularly interesting problem from a technical perspective: anatomically similar regions of the right and left breasts must be compared, but there are few reliable landmarks on which to base registration. The nipple is always present, and is usually visualized in profile, and the outline of the breast (skinline) may also be matched. However, the positioning of the breast at the time of mammography determines first the amount of tissue visualized at the chest wall boundary and second the extent of the breast outline. Features on the outline such as the inframammary fold are not always visualized [31]. Compression also has an impact on the appearance of the mammogram; this is an added complication when comparing images taken at different times. In the mediolateral oblique view, the line of the pectoral muscle and the position of the nipple can be used to define a frame of reference for alignment. In the craniocaudal view it is more difficult since the orientation of the chest wall boundary of the film depends entirely on patient positioning. Various methods have been investigated as means of comparison; bilateral subtraction, non-rigid registration and transportation have all been used in breast imaging [32–35]. An example of a multiresolution transportation-based approach to the detection of asymmetry is shown in Figure 3. Here, pixel values (which represent X-ray density) are moved from right to left breast images to make them similar. A cost is associated with distance and quantity, and an efficient solution is found. The output images represent cost associated with transportation to each destination pixel. Few asymmetry detection algorithms have been evaluated on data sets of significant size to enable comparison, and only one of the current CAD systems claims to detect asymmetry [29].

View larger version (60K): [in this window] [in a new window]   Figure 2. Craniocaudal screening mammograms. (a) Normal and (b) asymmetric (cancer).

View larger version (20K): [in this window] [in a new window]   Figure 3. Automated detection of asymmetry using transportation. (a) The right breast image; (b) the left breast image; (c) the pixels associated with high transportation costs (red and yellow) highlighted.

	Computer-based aids for film readers

Top Abstract Introduction Computer-based detection of... Computer-based aids for film... The potential of CAD... The clinical evaluation of... Conclusions References

A step back from this is automated pre-screening, in which the computer is used to sort cases into two categories: "abnormal or equivocal" and "normal". In the pre-screening model, the radiologist would view only those cases in which the computer detected something suspicious along with a small subset of the normal cases for quality control purposes [36]. The overall sensitivity with pre-screening is limited by the sensitivity of the computer system, as any cancer cases erroneously classified as normal by the computer would be unlikely to be detected. The gain in terms of reduced time spent interpreting the cases depends on the specificity of the computer system, but it is not enough to calculate the proportion of cases viewed: those deemed normal by the computer are likely to include the majority of mammograms of very fatty breasts which the human film reader can also dismiss very rapidly.

Over the last few years, detection algorithms have been used clinically as components of CAD systems, which aim to aid the film reader by drawing attention to suspicious regions of the original mammogram in a process known as prompting [37]. The aim of CAD is to ensure that all potentially significant regions of the mammogram are examined (to avoid errors caused by failure of the reader to adequately scan the whole mammogram), and given due consideration (to avoid errors in which abnormalities are detected but dismissed as being normal). Prompting systems require digital images, either acquired directly or obtained by digitizing film images. Algorithms are then applied to detect specific types of candidate abnormalities such as microcalcification clusters and masses. The most suspicious locations are marked in a prompt image, which is usually a low-resolution version of the mammogram. Before consulting the prompt images for a given case, the reader should make a thorough initial unprompted search of the original mammograms. This would ensure that the sensitivity of a reader with CAD at least equals their unprompted sensitivity. The reader then accesses the prompt images and re-evaluates the case, checking marked locations and noting any new findings. If the system is used as intended, and the algorithms are sufficiently sensitive and specific, the process should lead to an improvement in the reader's detection performance. The sensitivity of a reader using CAD should be limited neither by that reader's unprompted performance, nor by the sensitivity of the individual algorithms in the CAD system.

There are now a number of commercially available CAD systems, the first of which to be marketed extensively is the R2 ImageChecker (R2, CA, USA) [28] which detects and prompts potential masses and microcalcification clusters. This system incorporates a digitizer to convert film mammograms to digital format. Bar-coded case separators are used to relate each case to the corresponding prompt image. The prompts are presented on a small monitor positioned near the viewer on which the mammograms are displayed, with paper printouts of the prompt images available for back-up. The software provides information about the strength of evidence that caused each region to be prompted, and enables detailed examination of the prompted regions by means of magnified images accessible via a touch screen.

To date, three systems have been approved by the Food and Drug Administration (FDA), and insurance companies in the USA now reimburse an extra $17 per case when CAD is used. The MammoReader from iCAD (Instrumentarium Imaging Inc., Milwaukee, USA) [29] and CADx's Second Look [30] both operate on a similar principle to the ImageChecker, although the systems are based on different detection algorithms, and thus respond differently to potential abnormalities. There are also some practical differences in how the prompting information is displayed. These two systems are soon to be merged. The development of CAD is a fast growing field, and a number of other systems are at a stage where they will soon become contenders in the CAD market.

	The potential of CAD to aid screening

Top Abstract Introduction Computer-based detection of... Computer-based aids for film... The potential of CAD... The clinical evaluation of... Conclusions References

 
The aim of CAD is to improve detection performance by reducing the number of false negatives (missed cancers) [38, 39]. Some missed cancers are due to failure of the radiologist to search the whole of the mammogram adequately. These errors may be caused by a number of different factors including fatigue, distraction, viewing conditions and over-reliance on prior knowledge of the most probable locations of cancers. Missed cancers of this sort may have unambiguous signs of abnormality and may be avoided by double reading. A correctly placed CAD prompt should at least ensure that the reader searches the part of the image containing the cancer.

None of the commercial systems claims perfect sensitivity and specificity of their algorithms, and nor do they claim to detect all manifestations of cancer. The initial unprompted search is thus vital to the success of CAD. In addition to errors caused by failure to search the image thoroughly, cancers may also be missed if the signs are detected by the reader but wrongly dismissed as being normal. These cancers are more likely to be subtle in appearance, with similar features to normal background structures and benign abnormalities. In this case, a correctly placed prompt should add weight to the conviction of the reader that there is actually an abnormality present, thus reducing the possibility of misclassification. Many of the very early cancers which can be seen during retrospective analysis of screening films taken prior to cancer being detected show only subtle changes, but there is evidence that CAD systems are sensitive enough to prompt in such cases [2, 40].

The majority of four-film cases presented to current CAD systems will be prompted, regardless of whether or not they contain a cancer. In breast screening programmes, over 99% of women screened will have normal mammograms. With a false prompt rate of 0.5 per image, this would mean that only 1 in 100 prompts would actually correspond to a cancer; the remaining false positive prompts would have to be disregarded by the film reader. Some of these false prompts may be very easy to dismiss, for example if they mark clearly benign calcifications, crossing ducts or image artefacts. However, the overall effect of a high ratio of false prompts to true ones will be to reduce the weighting placed by the reader on any given prompt, thus reducing the potential of CAD to overcome misclassification errors. False prompts may also degrade performance by acting as distractors and drawing attention away from genuinely abnormal regions.

For algorithms which are very sensitive, such as microcalcification detection algorithms, there is a danger that the reader may become over-reliant on prompts, and miss those cancers that the system fails to detect. If CAD is being used as intended, the majority of these cancers should be detected in the initial, unprompted search. However, some studies have shown that readers took no longer to read with CAD than they did without it, which could indicate that they reduced their initial unprompted search of the image in the knowledge that they would be making a second search with the aid of prompts [41, 42].

There are many practical considerations associated with the introduction of CAD into a screening programme. In larger centres in the UK, the mammograms of approximately 200 women must be interpreted each day. This will include both standard and large format films, with any previous screening films displayed alongside current mammograms to enable comparison. Radiologists read up to 130 cases per hour [2], although the amount of time spent on each case is highly variable. In the NHSBSP, the women screened are between 50 years and 69 years and a large proportion of them have fatty breasts which are easily assessed. Mammograms showing fatty-glandular and dense breasts are more time-consuming to read. Clearly, the use of CAD will increase the time taken for an individual reader to review the films, since it is additional to the initial unprompted evaluation, but it is unlikely to be prohibitively slow in practice. The digitization and processing of films from larger centres is an issue which needs to be addressed; to cope with current workloads, two systems (or an overnight run) would be required. Medicolegal implications may also need to be addressed, for example in the case where a radiologist chooses to dismiss a prompt which is marking a genuine abnormality.

	The clinical evaluation of CAD

Top Abstract Introduction Computer-based detection of... Computer-based aids for film... The potential of CAD... The clinical evaluation of... Conclusions References

 
To date, much of the published information about human performance with CAD has been underpinned by the belief that if CAD shows a high prompt sensitivity on films containing cancers which are likely to be missed, then performance will be increased. In practice, the situation is likely to be more complex because of the adverse effects of false negative and false positive prompting.

The majority of published trials of CAD have been retrospective. The most significant prospective study took place over a 12 month period during which two readers examined screening mammograms unprompted, recorded their findings and then re-examined the films with the aid of prompts [43]. A 19.5% increase in the number of cancers detected was reported, with a 5% increase in the proportion of early cancers found at screening. These results are very encouraging; however with only two readers it is difficult to generalize the results to readers with a wider range of experience. Furthermore, they were reading with the aid of a "safety net"; they knew that their unprompted search was preliminary to a further search with prompts, and this could have adversely affected their baseline measure of unprompted performance.

One of the major limitations of many of the published retrospective studies is that many have chosen to load their test cases with cancers, rather than use a realistic screening mix in which fewer than 1% of cases show cancer [40, 42, 44–46]. This methodology has clearly been adopted for reasons of efficiency; it is very costly to have readers review the large numbers of cases required to evaluate reader sensitivity on a screening mix of cases. However, by loading the data they alter the ratio of false to true prompts. This artificially increases the readers' expectation that any given prompt will mark a cancer, leading to an overestimate of prompted performance and making generalization to the screening programme impossible.

Other studies have involved only a small number of readers, and are limited by the effects of natural differences in reading performance and practice between individuals. The degree of training (or in some cases, lack of it) may also have influenced results; we do not yet know how long it takes, or how many prompted cases must be read, for a reader to attain stable performance with CAD. Finally, since the spectrum of appearances of cancers is so wide, selection of small sets of cases is likely to lead to the results being over-dependent on the actual case-mix used.

Most published studies have shown improvement with CAD [43, 46]. One exception is a recent UK study of 50 readers interpreting 180 cases of which one third were cancers: no significant difference between prompted and unprompted conditions was found [42]. This result could, however, have been influenced by lack of reader training prior to the study, or to the artificial nature of the experiment.

	Conclusions

Top Abstract Introduction Computer-based detection of... Computer-based aids for film... The potential of CAD... The clinical evaluation of... Conclusions References

 
The detection of early breast cancer in mammograms is a challenge both for radiologists and for computer-based methods. Recent progress, particularly in the automated detection of microcalcification clusters and soft tissue masses, has enabled the development of CAD systems to aid radiologists by drawing their attention to potential abnormalities. Successful CAD requires algorithms that are both sensitive and specific, since false markers reduce the effectiveness of prompting. In the past few years there have been significant developments both in improving algorithm performance and in engineering clinical systems. Although there have been a number of clinical evaluations of CAD, it still remains to be seen whether systems with the current levels of algorithm performance can realise the potential of CAD in a screening programme in which over 99% of cases are normal. In the future, soft copy reading with the aid of CAD may provide an effective method of reading digital mammograms. Any reduction in spatial resolution of the original images would affect mainly the detection and analysis of microcalcifications, but these are tasks which can already be performed well by computer at similar resolutions; current algorithms can detect over 98% of clusters. A further advantage of linking CAD with soft copy reading is that the prompts can be presented in the same physical location as the abnormality concerned, avoiding unnecessary and distracting eye movements.

	Acknowledgments

 
I am grateful for the support and collaboration of colleagues in the NHSBSP, in particular Dr Caroline Boggis of the Nightingale Breast Centre, and for financial support from the NHSBSP, CRUK and EPSRC. The illustrations were provided by Michael Board.

	References

Top Abstract Introduction Computer-based detection of... Computer-based aids for film... The potential of CAD... The clinical evaluation of... Conclusions References

 

1. Elmore JG, Miglioretti DL, Carney PA. Does practice make perfect when interpreting mammography? Part II. J Natl Cancer Inst 2003;95:250–2.[Free Full Text]
2. Astley SM, Boggis CRM, Walker K, Wallace S, Tomkinson S, Hillier V, et al. An evaluation of a commercial prompting system in a busy screening centre. In: Peitgen H-O, editor. Proceedings of the Sixth International Workshop on Digital Mammography 2002, Berlin Heidelberg: Springer Verlag, 2002;471–5.
3. Huynh PT, Jarolimek AM, Daye S. The false negative mammogram. RadioGraphics 1998;18:1137–54.[Abstract]
4. Britton PD, McCann J, O'Driscoll D, Hunnam G, Warren RML. Interval cancer peer review in East Anglia: implications for monitoring doctors as well as the NHS Breast Screening Programme. Clin Radiol Vol 2000;56:44–9.
5. Blanks RG, Wallis MG, Moss SM. A comparison of cancer detection rates achieved by breast cancer screening programmes by number of readers, for one and two view mammography: results from the UK National Health Service breast screening Programme. J Med Screening 1998;5:195–201.[Abstract/Free Full Text]
6. Dinnes J, Moss S, Melia J, Blanks R, Song F, Kleijnen J. Effectiveness and cost-effectiveness of double reading of mammograms in breast cancer screening: findings of a systematic review. The Breast 2001;10:455–63.
7. Tarassenko L, Hayton P, Cerneaz N, Brady M. Novelty detection for the identification of masses in mammograms. Proc 4th International Conference on Artificial Neural Networks, Cambridge, UK. 1995;442–7.
8. Hoare SW, Asbridge D, Beatty PCW. On-line novelty detection for artefact identification in automatic anaesthesia record keeping. Med Engineering Phys 2002;24:673–81.[CrossRef]
9. Rose C, Taylor CJ. A statistical model of texture for medical image synthesis and analysis. In: Proc. Medical Image Understanding and Analysis. 2003:1–4.
10. Poissonier M, Brady JM. Noise equalization, film screen artifacts and density representation. In: Yaffe MJ, editor. Proceedings of the Fifth International Workshop on Digital Mammography. Wisconsin: Medical Physics Publishing 2001:605–11.
11. Rick A, Muller S, Bothorol S, Grimaud M. Quantitative modeling of microcalcification detection in digital mammography. In: Proceedings of Medical Image Computing and Computer-Assisted Intervention. 1999:32–41.
12. Hagihara Y, Kobatake H, Nawano S, Takeo H. Accurate detection of microcalcifications on mammograms by improvement of morphological processing. In: Yaffe MJ, editor. Proceedings of the Fifth International Workshop on Digital Mammography. Wisconsin: Medical Physics Publishing, 2001:193–7.
13. Chan H-P, Doi K, Galhotra S, Vyborny C, MacMahon H, Jokich P. Image feature analysis and computer-aided diagnosis in digital mammography. (I) Automated detection of microcalcifications in mammography. Med Physics 1987;14:538–48.[CrossRef][Medline]
14. Gurcan MN, Chan H-P, Sahiner B, Hadjiiski L, Petrick N, Helvie MA. Optimal neural network architecture selection: improvements in the computerized detection of microcalcifications. Academic Radiology 2002;9:420–9.[CrossRef][Medline]
15. Zhang X, Hara T, Fujita H, Shinohara N, Ooe Y, Iwase T, et al. Artificial neural network method for detecting clustered microcalcifications in masses on mammograms. In: Peitgen H-O, editor. Proceedings of the Sixth International Workshop on Digital Mammography. Berlin Heidelberg: Springer Verlag, 2002:391–3.
16. Karssemeijer N. Adaptive noise equalization and recognition of microcalcification clusters in mammograms. Int J Pattern Recogn Artificial Intell 1993;7:1357–76.[CrossRef]
17. Suckling J, Astley S, Betal D, Cerneaz N, Dance DR, Kok S-L, et al. The mammographic image analysis society digital mammogram database. Exerpta Medica. International Congress Series 1994;1069:375–8.
18. Heath M, Bowyer K, Kopans D, Moore R, Kegelmeyer P. The digital database for screening mammography. In: Yaffe MS, editor. Digital Mammography. Toronto, Canada: Medical Physics Publishing, 2001;457–60.
19. Alberdi E, Taylor P, Lee R, Fox J, Sordo M, Todd-Pokropek A. CADMIUM II: Acquisition and representation of radiological knowledge for computerized decision support in mammography. Proc AMIA Annual Symposium. 2000:7–11.
20. Kahn CE Jr, Roberts LM, Shaffer KA, Haddawy P. Construction of a Bayesian network for mammographic diagnosis of breast cancer. Computers Med Biol 1997;27:19–29.
21. Niklason LT, Christian BT, Niklason LE, Kopans DB, Castleberry DE, Opsahl-Ong BH, et al. Digital tomosynthesis in breast imaging. Radiology 1997;205:399–406.[Abstract]
22. Karssemeijer N, te Brake G. Detection of stellate distortions in mammograms. IEEE Trans Medical Imaging 1996;15:611–9.[CrossRef]
23. Zwiggelaar R, Parr T, Schumm J, Hutt I, Astley SM, Taylor CJ, et al. Model-based detection of spiculated lesions in mammograms. Medical Image Analysis 1999;3:39–62.[CrossRef][Medline]
24. Matsubara T, Ichikawa T, Hara T, Fujita H, Kasai S, Endo T, et al. Automated detection methods for architectural distortions around skinline and within mammary gland on mammograms. Computer-Assisted Radiology and Surgery: Proceedings of the 17th International Congress London, 2003;950–5.
25. Miller P, Astley S. Classification of breast tissue by texture analysis. Image and Vision Computing 1993;10:277–82.
26. Suckling J, Dance DR, Moskovic E, Lewis DJ, Blacker SG. Segmentation of mammograms using multiple linked self-organizing neural networks. Med Phys 1995;22:145–52.[CrossRef][Medline]
27. Marias K, Highnam R, Brady M, Parbhoo S, Seifalian A. Assessing the role of quantitative analysis of mammograms in describing breast density changes in women with HRT. In: Peitgen H-O, editor. Proceedings of the Sixth International Workshop on Digital Mammography. Berlin Heidelberg: Springer Verlag, 2002:547–51.
28. R2 ImageChecker (accessed 12.08.03) http://www.r2tech.com
29. iCAD MammoRead (accessed 12.08.03) http://www.imaging.com/mammo/issi.htm
30. CADx Second Look (accessed 12.08.03) http://www.icadmed.com
31. Caulkin S, Astley SM, Asquith J, Boggis C. Sites of occurrence of malignancies in mammograms. In: Karssemeijer N, Thijssen M, Hendriks J, van Erning L, editors, Digital Mammography, Kluwer Academic, Nijmegen, 1998.
32. Wirth MA, Narhan J, Gray D. Nonrigid mammogram registration using mutual information. In: Sonka M, Fitzpatrick JM, editors. Proceedings of SPIE Medical Imaging: Image Processing. 2002;4684:562–73.
33. Rueckert D, Sonoda LI, Hayes C, Hill DLG, Leach MO, Hawkes DJ. Non-rigid registration using free-form deformations: application to breast MR images. IEEE Transactions on Medical Imaging 1999;19:94–102.
34. Yin FF, Giger ML, Doi K, Vyborny CJ, Schmidt RA. Computerized detection of masses in digital mammograms: automated alignment of breast images and its effect on bilateral-subtraction technique. Med Phys 1994;21:445–52.[CrossRef][Medline]
35. Miller P, Astley SM. Automated detection of asymmetry in digitised mammograms. Int J Pattern Recogn Artificial Intell 1993;7:1461–76.[CrossRef]
36. Astley S, Mistry T, Boggis C, Hillier V. Should we use humans or a machine to pre-screen mammograms? In: Peitgen H-O, editor. Proceedings of the Sixth International Workshop on Digital Mammography. Berlin Heidelberg: Springer Verlag, 2002:476–80.
37. Computer-Aided Detection in Mammography: Working party of the Radiologists Quality Assurance Co-ordinating Group, NHSBSP Publications No 48, Jan 2001.
38. Gale AG, Savage CJ, Pawley EF, Wilson ARM, Roebuck EJ. Breast screening: visual search and observer performance. In: Kundel HL, editor. Image perception, Proc SPIE 1994;2166:66–75.
39. Blanks RG, Wallis MG, Given-Wilson R. Observer variability in cancer detection during routine repeat incident mammographic screening in a study of two versus one view mammography. J Med Screen 1999;6:152–8.[Abstract/Free Full Text]
40. Warren Burhenne LJ, Wood SA, D-Orsi CJ, Feig SA, Kopans DB, O'Shaughnessy KF, et al. Potential contribution of computer-aided detection to the sensitivity of screening mammography. Radiology 2000;215:554–62.[Abstract/Free Full Text]
41. Champness J, Taylor P, Given-Wilson R. Impact of computer-placed prompts on sensitivity and specificity with different groups of mammographic film readers. In: Peitgen H-O, editor. Proceedings of the Sixth International Workshop on Digital Mammography. Heidelberg: Springer Verlag, 2002:276–80.
42. Taylor PM, Champness J, Given-Wilson R, Potts HW, Johnston K. An evaluation of the impact of computer-based prompts on screen readers' interpretation of mammograms. Br J Radiol 2004;77:21–7.[Abstract/Free Full Text]
43. Freer TW, Ulissey MJ. Screening mammography with computer-aided detection: prospective study of 12,860 patients in a community breast center. Radiology 2001;220:781–6.[Abstract/Free Full Text]
44. Garvican L, Field S. A pilot evaluation of the R2 Image Checker System, Users' Response in the Detection of Interval Breast Cancers on Previous Screening Films. Clin Radiol 2001;56:833–7.[CrossRef][Medline]
45. Thurfjell E, Thurfjell MG, Egge E, Bjurstam N. Sensitivity and specificity of computer-assisted breast cancer detection in mammography screening. Acta Radiologica 1998;39:384–8.[Medline]
46. Karssemeijer N, Otten JDM, Verbeek ALM, Groenewoud JH, de Koning HJ, Hendriks JHC, et al. Computer-aided detection versus independent double reading of masses on mammograms. Radiology 2003;227:192–200.[Abstract/Free Full Text]
47. Board M, Astley S. A new method for evaluating and optimising mammographic detection algorithms. In: Peitgen H-O, editor. Proceedings of the Sixth International Workshop on Digital Mammography. Heidelberg: Springer Verlag, 2002:257–61.

This article has been cited by other articles:

				D. Georgian-Smith, R. H. Moore, E. Halpern, E. D. Yeh, E. A. Rafferty, H. A. D'Alessandro, M. Staffa, D. A. Hall, K. A. McCarthy, and D. B. Kopans Blinded Comparison of Computer-Aided Detection with Human Second Reading in Screening Mammography Am. J. Roentgenol., November 1, 2007; 189(5): 1135 - 1141. [Abstract] [Full Text] [PDF]

				N Karssemeijer, J D M Otten, H Rijken, and R Holland Computer aided detection of masses in mammograms as decision support Br. J. Radiol., December 1, 2006; 79(Special_Issue_2): S123 - S126. [Abstract] [Full Text] [PDF]

This Article Abstract Figures Only Full Text (PDF) Services Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Astley, S M Search for Related Content PubMed PubMed Citation Articles by Astley, S M